Lightweight And High-precision Facial Landmark Detection Algorithm Based On Convolutional Neural Network

As one of the important research directions in the field of computer vision,facial landmark detection is applied to many visual tasks such as face edit,face pose estimation,face reconstruction,fatigue driving detection and face recognition.However,the existing facial landmark detection algorithm can't take into account both detection accuracy and network lightweight.In this thesis,we designed a lightweight facial landmark detection network named LFLDNet,which based on convolutional neural networks,and introduced the idea of knowledge distillation,using a complex teacher network to guide LFLDNet for training,improving the detection accuracy of LFLDNet;At the same time,a pixel balance adaptive wing loss function is proposed,which overcomes the problem that the widely used loss function is insensitive to small errors and large errors,and ignores the difference in the number of foreground and background pixels,which limit detection accuracy of the algorithm;Further based on the above methods,While realizing the facial landmark detection algorithm is lightweight,it has effectively improved the detection accuracy of the algorithm.The main research work is as follows:1.According to the problem that the current facial landmark detection algorithm based on convolutional neural network has too many network parameters,which was not conducive to deployment on embedded devices with limited memory.a lightweight facial landmark detection network was designed which named LFLDNet.By using a lightweight and efficient MobileNetV3 as the encoder and using group deconvolution as the decoder,the amount of network parameters was reduced significantly.2.A detection accuracy improvement method combining pixel-wise distillation and feature similarity distillation was proposed.By aligning the predict heatmap of the teacher network and LFLDNet and the feature similarity matrix of the intermediate feature maps of the two networks,the knowledge learned by the teacher network Distill into LFLDNet,which improving the detection accuracy of LFLDNet.Comparative experiments were conducted on two mainstream datasets 300 W and WFLW,and the experiments data demonstrated that the normalized mean error of the proposed method in the two datasets was reduced by 3.91% and 3.15% respectively,compared to the method without knowledge distillation.3.The widely used MSE loss function is insensitive to small errors,and L1 loss function is insensitive to large errors,and both of them ignores the difference between the number of foreground and background pixels,which limit the detection accuracy of the algorithm.To solve this problem,a pixel balance adaptive wing loss function is proposed.It adaptively adjusts the optimization method according to the value of the error,which is more sensitive to small errors and large errors.Meanwhile,it realizes the foreground by increasing the weight of the foreground pixel in the loss function and reducing the weight of the background pixel in the loss function,which achieves a number balance between foreground and background pixels.Because of these attributes,this loss function obviously improves the performance of the facial landmark detection algorithm.Comparative experiments were conducted on two mainstream datasets of 300 W and WFLW.The experimental data demonstrated that the normalization mean error of the proposed loss function in the two datasets is reduced by 8.04 % and 9.94% compared to the L1 loss function,and reduced by 6.98%and 10.14% in comparison to the MSE loss function.